| title |
aliases |
tags |
sources |
created |
updated |
| The .claude Folder — Full Configuration Reference |
| dot-claude |
| claude-folder |
| claude-configuration |
|
| claude-code |
| configuration |
| hooks |
| skills |
| agents |
| plugins |
| mcp |
|
|
2026-04-29 |
2026-04-29 |
The .claude Folder — Full Configuration Reference
"If you're using Claude Code without setting up .claude folder, you're using 10% of it."
The .claude folder is where Claude Code's real power lives. Most users never open it.
Mental Model
Three tiers of execution:
| Tier |
Example |
Behavior |
| Advisory |
CLAUDE.md |
Claude reads and follows — but can be overridden |
| Deterministic |
.claude/hooks/ |
Fires every time, no exceptions |
| On-demand |
.claude/skills/ |
Loaded only when the task matches |
Files and Folders
CLAUDE.md (project root)
- Your project rulebook — coding style, conventions, gotchas
- Claude reads it before every task; stops you repeating yourself in every prompt
- Advisory: Claude follows it, not enforced at the system level
.claude/hooks/
- Shell scripts that fire on lifecycle events (session start, after edits, before compaction, etc.)
- Deterministic — Claude cannot skip them
- Common uses: auto-commit after edits, load project context on startup, save state before compaction
.claude/skills/
- Drop a
SKILL.md inside any folder; Claude invokes it when the task matches
- Skills are scoped — no skill bleeds into unrelated work
- Examples: carousel generation skill, pacing drills skill
.claude/agents/
- Subagents with isolated context windows
- Run in parallel; report back to the main thread
- Common patterns: code reviewer, researcher, log analyzer
.claude/commands/
- Slash commands (legacy, still functional)
- Example:
/ship → lint + build + deploy in one go
.claude/plugins/
- First-class in 2026
- Bundle hooks + skills + agents + MCP servers into one shareable package
- The "real unlock" for team sharing and distribution
.claude/rules/
- Path-scoped rules loaded only when files match a glob pattern
- Prevents API rules from polluting frontend context (and vice versa)
.mcp.json (project root, not inside .claude/)
- MCP server config for the project
- Claude connects them automatically on session start
Key Takeaways
.claude folder = Claude Code's brain — ignoring it means ignoring most of the product
- CLAUDE.md is advisory; hooks are deterministic; skills load on demand — this is the core mental model
- Hooks are the enforcement layer — use them for anything that must always happen
- Skills scope Claude's capabilities to context — prevents wrong tools firing on wrong tasks
- Agents keep the main context clean — parallel subagents handle heavy lifting
- Plugins (2026) are the distribution primitive — bundle everything into one installable unit
.mcp.json lives at the project root, not inside .claude/
Related Articles
Sources